Separate graph TIDs from backend UIDs in JSON repros

[hwanseoc]

@coderabbitai ignore

Why

Graph JSON v1 uses tensor names as references. Duplicate and anonymous names can collapse distinct tensors and make extracted repros incorrect.

This draft is the hybrid-identity alternative to #280:

• tid: graph-local structural identity used by graph JSON and repro extraction
• uid: user/backend identity used by variant packs and execution plans

Architecture

• emit graph JSON 2.0
• assign graph-local TIDs at tensor creation
• reference node ports and ragged offsets by TID
• preserve UID and whether it was explicitly assigned
• emit both TID and UID in tensor dumps; join repro dumps by TID
• use process-local gid to correlate serialized graphs with executions
• leave backend UID assignment behavior unchanged

Behavior shared with #280

• tensor-list graph schema with optional names
• typed rejection of malformed versions, tensor tables, duplicate identities, and dangling references
• versioned plan-only serialization without gid or graph structure
• per-execution repro correlation
• simple, FP8, and MXFP8 SDPA extraction, including sequence lengths and ragged offsets
• migrated XLA serialization sample and native lowered-JSON consumers
• canonicalization of shared ragged-offset objects during deserialization

Intentional difference from #280

TID and UID are explicit, non-interchangeable namespaces. Structural serialization does not assign missing backend UIDs. #280 instead uses UID for both structure and execution identity.

Compatibility

• Structural graph JSON v1 is rejected; no ambiguous upgrade is attempted.
• Plan/backend consumers remain UID-based.
• gid is process-local execution correlation, not persistent identity.

Dependencies

#386 contains the architecture-independent runtime tensor-dump and FP8/MXFP8 tester fixes.

Verification

Based on develop 526020e; B200, CUDA 13.1, cuDNN backend 9.30:

• C++ graph/serialization suite: 75 assertions in 10 cases passed
• chained two-operation TID allocation torture test: 11 assertions passed
• shared ragged-offset canonicalization: 5 assertions passed
• native lowered-JSON regression: 1 passed
• deviceless deserialize API-contract regression: 1 passed
• complete repro-tool suite: 71 passed, 2 MXFP8 target skips
• live parity against Use UID-based graph JSON v2 for repro extraction #280 after normalizing only UID/TID identity: forward, ragged forward, backward, and FP8 payloads match

Draft question

Should graph-local TID allocation remain at tensor creation sites, or move to a centralized graph-owned allocator before merge?

[hwanseoc]

@cudnn-ci-bot run

[cudnn-ci-bot]

🚀 Running mirror pipeline

Branch: cudnn-gh/pr-383-a7b38d3
Pipeline: 57833052